Outline

Our primary question is: What proportion of people with diarrhea seek care at a place that they’d likely get counted (i.e., hospital/clinic) when they have diarrhea?

  1. Look at care seeking in the raw data by:

    • Symptoms/case definition
    • Sampling method
    • Study type
    • Urban/rural
    • How study question is phrased (multiple choice, timing of care seeking, recall period)
    • Study population/respondent
    • Outbreak setting
    • Income group
    • Geographic region
    • Age
  2. Estimate proportion seeking care at hospital or clinic overall, following methods from Wiens et al PLOS Med

  3. Compare with results if we were to run a RE meta-analysis using default settings in metafor/meta package

  4. Examine differences by age in studies with age stratifications

  5. Examine factors associated with care seeking at a hospital or clinical

Our secondary question is: Where else do they seek care?

  1. Look at care seeking in the raw data by variables that came out as key in main analysis:

    • Symptoms/case definition
    • Income group
    • Timing of care seeking
    • Age


Systematic review

Fig 1. PRISMA diagram

PRISMA diagram.
PRISMA diagram.


Data

Note: Unless otherwise noted, the data presented below represents 161 studies (381 stratified observation entries) in the primary, non-overlapping dataset that asked study participants whether or not the DID seek care for a past diarrheal illness.

We also have information on whether individuals would seek care for a hypothetical diarrheal illness from 27 studies (96 stratified observation entries).

Data coverage by geography

Number of stratified observation entries in the primary dataset at each administrative level by country. Countries with >10 observations displayed as 10.

variable observations
admin0 27
admin1 80
admin2 157
admin3 117
region observations
Middle Africa 8
Middle East & North Africa 8
Southern Africa 9
Europe & Central Asia 10
North America 12
Western Africa 36
East Asia & Pacific 42
Latin America & Caribbean 72
South Asia 84
Eastern Africa 100


Data coverage over time

Number of stratified observation entries in the primary dataset within different time periods. Year represents the year sampling was completed. Excludes 6 entries missing study dates.

variable observations
2000~2004 30
2005~2009 113
2010~2014 157
2015~2022 74
NA 7


Study (entry_id) characteristics

variable value Total Percent
Study type Survey 304 79.8
Intervention 50 13.1
Surveillance 23 6.0
Case control 4 1.0
Sampling method Cluster 142 37.3
Simple random 101 26.5
All cases or households 53 13.9
Stratified 37 9.7
Systematic 37 9.7
Other (describe in notes) 11 2.9
Study population Caregiver 212 55.6
Head of household 67 17.6
Resident 51 13.4
Resident with telephone/email access 21 5.5
Caregiver of patient at health facility 11 2.9
Patient at health facility 6 1.6
Relative/Caregiver (of adult) 6 1.6
Child with malnutrition 2 0.5
Other 2 0.5
Orphan 1 0.3
Street dweller 1 0.3
University student 1 0.3
Outbreak described No 308 80.8
Yes 73 19.2
Location desc Rural 180 47.2
Urban 99 26.0
Urban and Rural 73 19.2
Peri-Urban 24 6.3
IDP or Refugee Camp 4 1.0
Urban and Peri-Urban 1 0.3
Income group Lower middle income 208 54.6
Upper middle income 73 19.2
Low income 69 18.1
High income 31 8.1
Case definition Diarrhea 264 69.3
Severe diarrhea or cholera 75 19.7
Gastroenteritis or non-Vc etiologies 42 11.0


Care seeking questions

Note: “Other” regarding self or child almost always refers to all household members, or is a combination of self and child that we could not disaggregate.

variable value Total Percent
Time care Any care 320 84.0
First source 50 13.1
Prior to current visit 9 2.4
Within 24 h 2 0.5
Self or child Child 253 66.4
Other 105 27.6
Self 23 6.0
Recall time 14-15 days 172 45.1
27-30 days 90 23.6
170-365 days 55 14.4
Not reported 43 11.3
2-7 days 13 3.4
42-90 days 8 2.1
Mult choice No 337 88.5
Yes 44 11.5


Proportion seeking care at a hospital or clinic

Note: This excludes individuals seeking care exclusively or explicitly at private hospitals/clinics, i.e., this includes location categories “Hospital/Clinic” (general) and “Public Hospital/Clinic” (more specific).

For multiple choice questions where there is more than one category of hospital/clinic, we take the largest proportion that sought care across them. This is not entirely accurate but the best we can do here. May want to return to this.


By study (entry_id) characteristics


By type of care seeking questions


Care seeking at hospital/clinic by age

Relationship between care seeking and proportion under 5 years

Overall


By case definition


By income group


By case definition and income group


Studies with either only kids or only adults


Meta-analysis for LMICs

Our goal is to estimate the proportion of individuals that seek care when they or their child have diarrhea symptoms in LMICs. Since diarrhea symptoms vary, and some of the categories above had very few observations (e.g., cholera and “other”) we have grouped these into three broad categories of case definitions (below). We’ll start by testing these along with other potential covariates in univariate analyses.

  1. Diarrhea (3+ loose stools in past 24 hours OR general/no explicit definition provided)
  2. Cholera (acute watery diarrhea of any severity, including resulting in death) or severe diarrhea (danger signs, dehydration, hospitalization, 1+ week duration, or death)
  3. Gastroenteritis or other etiologies (diarrhea or vomiting, including specific etiologies eg rotavirus or e coli)

We’ll build generalized linear models with a study-level random intercept following the approach of Wiens et al (2023) PLOS Medicine:

  1. Unadjusted

  2. Univariate models adjusted individually for:

    1. Study methodology and population:
      • case definition
      • timing of care seeking
      • regarding self or child
      • recall period
      • multiple choice questions
      • study population
    2. Contextual factors:
      • outbreak
      • urban/rural
  3. Multivatiate model(s) adjusted for factors identified as significant but not confounding/collinear in univariate models.

NB: Here our observations represent data aggregated to study_id, country, and potential covariate stratifications, for a total of 146 unique observations corresponding to 114 studies. For studies with multiple choice questions, we take the maximum sample size and number seeking care for studies where there were multiple categories of “hospital/clinic” as possible answers.

Prior predictive checks

For the primary analysis, we chose priors that matched the distribution of the data. Specifically, we used a Normal(-0.6,1.5) prior on alpha, where the mean of the distribution (-0.67) matches the median of the observation data (0.339) in logit space. The resulting prior predictive distribution has a median of 0.34 and a mean of 0.37.

In sensitivity analyses, we shift the mean of the prior distribution on alpha up or right 20% in probability space (i.e., mean of -0.38 for the normal distribution in logit space, median of 0.41 for the observation data) and down or left 20% (i.e., mean of -0.99 for the normal distribution in logit space, median of 0.27 for the observation data).


Unadjusted estimates

Bayesian approach

Posterior distributions of the proportion of people seeking care


Estimated proportion that seek care
Version Proportion (%)
Unadjusted 39.7 (9.5 - 77.4)
Unadjusted - shift prior left 37.1 (8 - 74.5)
Unadjusted - shift prior right 42.6 (11.1 - 80)


RE meta-anaysis a la ‘meta’

## Number of studies: k = 146
## Number of observations: o = 136822
## Number of events: e = 42388
## 
##                      proportion           95%-CI
## Random effects model     0.3455 [0.2941; 0.4007]
## 
## Quantifying heterogeneity:
##  tau^2 = 2.0548; tau = 1.4335; I^2 = 99.3% [99.2%; 99.3%]; H = 11.56 [11.26; 11.88]
## 
## Test of heterogeneity:
##              Q d.f. p-value
##  Wald 19392.96  145       0
##  LRT  32811.27  145       0
## 
## Details on meta-analytical method:
## - Random intercept logistic regression model
## - Maximum-likelihood estimator for tau^2
## - Logit transformation
## - Continuity correction of 0.5 in studies with zero cell frequencies
##   (only used to calculate individual study results)


Sub-analyses by age

Studies that are either only adults or only children

Bayesian approach

Odds of seeking care for diarrhea for adults vs. children, not adjusting for any other variables. We have 130 observations total here.

Variable Category Odds ratio
prop_five 0 1 [Reference]
1 1.07 (0.51 - 2.07)


RE meta-analysis a la ‘meta’
## Number of studies: k = 119
## Number of observations: o = 112825
## Number of events: e = 32724
## 
##                      proportion           95%-CI
## Random effects model     0.3432 [0.2877; 0.4034]
## 
## Quantifying heterogeneity:
##  tau^2 = 1.9896; tau = 1.4105; I^2 = 99.2% [99.1%; 99.2%]; H = 11.04 [10.71; 11.39]
## 
## Test of heterogeneity:
##              Q d.f. p-value
##  Wald 14391.78  118       0
##  LRT  23792.53  118       0
## 
## Results for subgroups (random effects model):
##                 k proportion           95%-CI  tau^2    tau        Q   I^2
## prop_five = 1  99     0.3436 [0.2842; 0.4084] 1.9190 1.3853 12552.51 99.2%
## prop_five = 0  20     0.3401 [0.2047; 0.5078] 2.3604 1.5363  1693.16 98.9%
## 
## Test for subgroup differences (random effects model):
##                   Q d.f. p-value
## Between groups 0.00    1  0.9671
## 
## Details on meta-analytical method:
## - Random intercept logistic regression model
## - Maximum-likelihood estimator for tau^2
## - Logit transformation


Studies with care seeking stratified by age

Bayesian approach

Odds of seeking care for diarrhea for adults vs. children, not adjusting for any other variables. We have data from 7 studies that report healthcare seeking at a hospital/clinic by age groups under/over five for LMICs.

Variable Category Odds ratio
prop_five 0 1 [Reference]
1 1.75 (0.29 - 6.04)


RE meta-analysis a la ‘meta’
## Number of studies: k = 14
## Number of observations: o = 26954
## Number of events: e = 7955
## 
##                      proportion           95%-CI
## Random effects model     0.3165 [0.1729; 0.5062]
## 
## Quantifying heterogeneity:
##  tau^2 = 2.0926; tau = 1.4466; I^2 = 99.2% [99.0%; 99.3%]; H = 10.96 [9.99; 12.02]
## 
## Test of heterogeneity:
##             Q d.f. p-value
##  Wald 1561.41   13       0
##  LRT  1775.73   13       0
## 
## Results for subgroups (random effects model):
##                 k proportion           95%-CI  tau^2    tau       Q   I^2
## prop_five = 0   7     0.2844 [0.1079; 0.5665] 2.4459 1.5639 1233.24 99.5%
## prop_five = 1   7     0.3461 [0.1583; 0.5984] 1.6565 1.2870  255.83 97.7%
## 
## Test for subgroup differences (random effects model):
##                   Q d.f. p-value
## Between groups 0.13    1  0.7218
## 
## Details on meta-analytical method:
## - Random intercept logistic regression model
## - Maximum-likelihood estimator for tau^2
## - Logit transformation


Univariate analyses

Odds of seeking care for diarrhea for each indicated variable, not adjusting for any other variables.

Variable Category Odds ratio
case_cat Diarrhea 1 [Reference]
Gastroenteritis or non-Vc etiologies 2.46 (0.77 - 6.1)
Severe diarrhea or cholera 3.21 (1.43 - 6.3) **
location_desc Non-urban 1 [Reference]
Urban 1.13 (0.64 - 1.8)
Urban and non-urban 1.53 (0.7 - 2.82)
mult_choice 0 1 [Reference]
1 1.1 (0.54 - 1.96)
outbreak_desc 0 1 [Reference]
1 3.14 (1.2 - 7.18) **
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 8.24 (2.19 - 21.4) **
Head of household 1.05 (0.54 - 1.9)
Other resident 1.27 (0.66 - 2.16)
recall_time over30_orNA 1 [Reference]
under30 0.38 (0.21 - 0.62) **
self_or_child Child 1 [Reference]
Other 0.86 (0.47 - 1.43)
Self 0.94 (0.26 - 2.53)
time_care Any care 1 [Reference]
First source 0.92 (0.44 - 1.69)
Prior to current visit 1.52 (0.17 - 5.88)


Tests for potential confounding

In analyses below, the dependent variable is listed in the header.

self_or_child (self/adult vs child)

Variable Category Odds ratio
case_cat Diarrhea 1 [Reference]
Gastroenteritis or non-Vc etiologies 4.55 (0.02 - 32.93)
Severe diarrhea or cholera 0.75 (0.01 - 4.29)
mult_choice 0 1 [Reference]
1 0.38 (0.01 - 1.94)
outbreak_desc 0 1 [Reference]
1 0.54 (0.01 - 3.09)
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 1.98 (0.01 - 13.03)
Head of household 1.94 (0.02 - 12.01)
Other resident 75.38 (9 - 298.64) **
recall_time over30_orNA 1 [Reference]
under30 0.2 (0.03 - 0.73) **
time_care Any care 1 [Reference]
First source 0.35 (0.01 - 1.75)
Prior to current visit 2.3 (0.01 - 14.42)


outbreak_desc

Variable Category Odds ratio
case_cat Diarrhea 1 [Reference]
Gastroenteritis or non-Vc etiologies 1.99 (0.02 - 12.6)
Severe diarrhea or cholera 24.95 (4.01 - 89.69) **
mult_choice 0 1 [Reference]
1 1.77 (0.16 - 7.49)
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 1.43 (0.01 - 8.62)
Head of household 5.8 (0.8 - 21.63)
Other resident 1.63 (0.15 - 6.46)
recall_time over30_orNA 1 [Reference]
under30 0.09 (0.01 - 0.27) **
self_or_child Child 1 [Reference]
Other 2.19 (0.34 - 7.68)
Self 1.61 (0.02 - 9.59)
time_care Any care 1 [Reference]
First source 0.49 (0.03 - 1.9)
Prior to current visit 2.24 (0.02 - 14.44)


case_cat

Severe diarrhea or cholera
Variable Category Odds ratio
mult_choice 0 1 [Reference]
1 1.46 (0.13 - 6.14)
outbreak_desc 0 1 [Reference]
1 26.42 (3.49 - 97.28) **
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 5.34 (0.14 - 28.6)
Head of household 4.76 (0.55 - 17.67)
Other resident 4.57 (0.56 - 17.1)
recall_time over30_orNA 1 [Reference]
under30 0.03 (0.01 - 0.09) **
self_or_child Child 1 [Reference]
Other 4.16 (0.72 - 13.85)
Self 1.4 (0.01 - 9.09)
time_care Any care 1 [Reference]
First source 1.12 (0.14 - 3.94)
Prior to current visit 9.8 (0.18 - 55.14)


Gastroenteritis or non-V cholerae
Variable Category Odds ratio
mult_choice 0 1 [Reference]
1 0.86 (0.04 - 3.91)
outbreak_desc 0 1 [Reference]
1 1.17 (0.01 - 7.25)
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 10.2 (0.28 - 53.83)
Head of household 0.61 (0.01 - 3.19)
Other resident 19.33 (2.81 - 68.96) **
recall_time over30_orNA 1 [Reference]
under30 0.17 (0.02 - 0.64) **
self_or_child Child 1 [Reference]
Other 12.9 (1.78 - 47.31) **
Self 2.16 (0.02 - 14.49)
time_care Any care 1 [Reference]
First source 0.61 (0.04 - 2.47)
Prior to current visit 3.2 (0.02 - 21.03)


Sub-analyses

Vc or severe diarrhea

Is there still an effect of being in an outbreak when we subset the data to just case definitions specific to cholera or severe diarrhea (including death)?

Variable Category Odds ratio
outbreak_desc 0 1 [Reference]
1 1.99 (0.35 - 6.14)


Diarrhea

Is there still an effect of recall period when we subset the data to just case definitions for general diarrhea (not severe or resulting in death)?

Variable Category Odds ratio
recall_time over30_orNA 1 [Reference]
under30 0.67 (0.33 - 1.21)


Multivariate analysis

In the above analyses, found that we cannot separate effects of:

  1. time_care, self_or_child, & case_cat (gastroenteritis or other etiology)
  2. recall_time, outbreak_desc, & case_cat (V. cholerae or severe)

The trends with self_or_child were also not significant or in the direction we expected, and so we’re focusing on methods and case_cat in the multivariate analyses.

We run two analyses below, adjusting for methods and then adjusting for methods and case_cat. We do not adjust for recall period in the case_cat analysis because there was potential collinearity/confounding between those variables.

  1. In the first model, we test case_cat adjusting for time_care, assuming that case definition or symptoms are more directly related to care seeking than outbreak context or recall period (seems reasonable but can discuss).
  2. In the second model, we only include methods variables: timing of care seeking.


Factors associated with variation in care seeking rates

Odds that an individual with diarrhea seeks care for themselves or their child overall or by alternate case definitions, adjusting for different ways that the care seeking questions were asked (i.e., timing of care itself).

Model including study population
Variable Category Odds ratio
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 8.38 (2.34 - 22.13) **
Head of household 1.06 (0.52 - 1.91)
Other resident 1.31 (0.64 - 2.23)
Model including study population and case definition
Variable Category Odds ratio
case_cat Diarrhea 1 [Reference]
Gastroenteritis or non-Vc etiologies 2.2 (0.66 - 5.49)
Severe diarrhea or cholera 2.94 (1.33 - 5.61) **
pop_cat Caregiver 1 [Reference]
Caregiver of patient at health facility 7 (1.85 - 18.82) **
Head of household 0.95 (0.48 - 1.75)
Other resident 0.99 (0.51 - 1.77)


Stratified estimates of care seeking by category

Stratifications correspond to questions that refer to any care seeking.

Version Model Variable Proportion (%)
Main result 1 Unadjusted 39.7 (9.5 - 77.4)
2 Diarrhea 38.6 (8.7 - 77)
2 Gastroenteritis or non-Vc etiologies 69.9 (25.2 - 98.8)
2 Severe diarrhea or cholera 79.1 (38.1 - 99.1)
3 Adjusted for methods 39 (8.7 - 77.1)
Shift prior left 1 Unadjusted 37.1 (8 - 74.5)
2 Diarrhea 35.6 (7.2 - 74.3)
2 Gastroenteritis or non-Vc etiologies 67.2 (21.4 - 98.8)
2 Severe diarrhea or cholera 77.4 (33.8 - 99.2)
3 Adjusted for methods 35.7 (7.1 - 74.4)
Shift prior right 1 Unadjusted 42.6 (11.1 - 80)
2 Diarrhea 41.2 (9.4 - 79.9)
2 Gastroenteritis or non-Vc etiologies 71.6 (26.7 - 98.7)
2 Severe diarrhea or cholera 80.6 (39.4 - 99.2)
3 Adjusted for methods 42.2 (10 - 80.3)



Forest plots

The “estimated” results are the study-level props from the model that includes time_care and case_cat.

## I2: 99.98747, 99.98791, 99.98834, ; tau2: 0.76853, 0.79656, 0.82605,


Where else do people seek care in LMICs?

Data

Overall proportion seeking care by source

By source and study characteristic

Case definition

Income group

Timing of care seeking

Self or child


Summary of key results

3) Found high variation between studies, resulting in wide uncertainty around global estimates of proportion seeking care at hospitals/clinics
  • Wide range in study-level estimates and global estimate credible intervals, very high I2
  • Estimated proportion seeking care varied between 4 modeling approaches (3 different priors for Bayesian GLM, and non-Bayesian RE meta-analysis): 35% to 51%.
4) The only factor that was clearly and consistently associated with variation in care seeking at hospitals/clinics was diarrhea case definition:
  • In univariate analysis, found that case definition, being in an outbreak, and being a caregiver of a patient at a health facility (as oppose to at home) were associated with higher odds of seeking care, while having a longer recall period was associated with lower odds of seeking care.
  • In bivariate analysis, found that we could not separate the effects of 1) recall period, case definition, and whether the study was about an individual or their child, or 2) recall period, case definition, and whether the study took place during an outbreak.
  • In sub-analysis, there was no longer an effect of being in an outbreak when sub-setting to just severe diarrhea or cholera, and there was no longer an effect of recall period when sub-setting to just general diarrhea.
  • In multivariate analysis, found that odds of seeking care for severe diarrhea or cholera were significantly higher than for general diarrhea, OR 3.15 (95% CrI 1.41-6.09), adjusting for timing of care seeking.

Additional things to consider

Sensitivity analyses (Skye found no differences):

  • study quality stratifications based on critical appraisals

Additional country- or study-level covariates to explore (Marissa working on):

  • Policies on free or affordable healthcare
  • Ratio of medical practitioners or CHWs to population

Supplementary analyses (potential, in rough order of importance):

  • would you seek care
  • assumptions about single-choice answers and where else people may seek care
  • recalculate proportion seeking care in hospital data to account for the fact that they’re already at hospital (use Bayes rule)